Dubstrata grounds market intelligence in a single graph and vector engine

Dubstrata monitors institutional risk and traces what causes market moves. It needed semantic search and causal graph traversal in the same place. With ArcadeDB, one engine does both, and hybrid retrievals typically finish in 20 to 30 milliseconds on a single cloud VM.

About Dubstrata

Dubstrata is an institutional risk monitoring and causal intelligence engine. It ingests continuous, high-volume market signals, macroeconomic disclosures and prediction market data, and turns fragmented news into point-in-time causal intelligence. Its customers are quantitative hedge funds, systematic trading desks, asset managers and enterprise risk teams. Dubstrata offers an API with real-time streaming feeds and is developing an MCP server for agentic workflows.

The challenge

Standard LLM and retrieval-augmented generation pipelines struggle with complex market events. They hallucinate, they lose track of when something happened, and they have no structural model of what causes what. Forecasting models face a related problem: without the full context of what moves the price of an asset, their predictions are weak.

Dubstrata's answer is to ground everything it delivers in a verified knowledge graph. Quantitative teams can then trace multi-hop causal chains and check factual consistency before they make a trading or risk decision. That requires semantic search to find the right starting points and graph traversal to follow the relationships from there, both in real time.

Why ArcadeDB

Dubstrata evaluated Neo4j, Memgraph, and dual-stack setups that run a dedicated graph database next to a separate vector database. ArcadeDB won for two reasons.

First, ArcadeDB combines graph traversal and vector search in one engine. That removed the work and risk of keeping two databases consistent with each other. Second, ArcadeDB is open source, so the team could start immediately and test quickly, without getting into prohibitive enterprise licensing.

“ArcadeDB is the most obvious alternative to Neo4j. It gives you graph traversal, vector search, and genuine ease of use in a single engine.”

Muzi Khuzwayo, Founder & CEO, Dubstrata

How Dubstrata uses ArcadeDB

For Dubstrata, the graph is the core truth layer. Vector search alone returns semantically similar items but has no awareness of how they relate. ArcadeDB lets the inference engine anchor on semantic vectors and then immediately traverse causal paths, entity networks and cross-market dependencies.

A native SQL vector query over 3,072-dimensional embeddings, backed by ArcadeDB's LSM vector indexes, finds the initial semantic anchors. Multi-hop Cypher traversals then extract the surrounding sub-graph of contextual causal paths for the inference engine. New events and context fragments arrive continuously and are written in batches, while the same graph serves real-time queries.

Results

On a single 8 vCPU, 32 GB virtual machine, a graph of more than 600,000 nodes and 1,000,000 edges serves hybrid vector and graph retrievals in 20 to 30 milliseconds at typical load. The 95th percentile stays under 50 milliseconds.

A lean deployment

ArcadeDB runs as a single Docker server on a standard Google Cloud VM with dedicated SSD volumes. Dubstrata describes it as a lean, cost-efficient cloud setup. One engine replaced what would otherwise have been two databases and the synchronization between them.

What comes next

As the graph grows, Dubstrata plans to move ArcadeDB to a dedicated compute server and then to a replicated setup. Based on its experience so far, the team expects the move to be straightforward.